Meta introduces Muse Ember, claiming lower AI task costs than rivals
The model matched Claude Haiku 5.5 on an independent intelligence benchmark, but Haiku performed better on a separate coding test.
Meta introduced Muse Ember on November 9, 2026, targeting recurring AI agent tasks such as request routing and document processing. The company describes Ember as its smallest, lowest-cost model and says it offers better performance per dollar than competing models.
Access is through Muse Code, Meta’s coding tool, and the Meta Model API. For each million tokens, Ember’s charges are US$0.10 for input and US$0.50 for output. Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna carry identical list prices.
Meta bases its task-cost comparison on the Intelligence Index, produced by independent benchmarking firm Artificial Analysis. At their highest settings, Haiku 5.5 and Ember both recorded 36; Luna recorded 28. Meta’s figures put the cost for each benchmark task at about US$0.14 using Ember, US$0.41 using Haiku and US$0.15 using Luna.
Other results varied by test. Ember achieved 37.8% on AutomationBench, which assesses complete business workflows; Luna registered 20.2% and Haiku 13.9%. On FrontierCode 2.0, which assesses whether code changes would be accepted by maintainers, Haiku’s 34.9% exceeded Ember’s 27.2%. Some additional results came from Meta’s internal testing.
Ember supports audio, video, images and text, accepting up to a million tokens in one input. Meta proposes pairing it with Muse Wonder, assigning planning to the larger model and high-volume work to Ember.
The company said its safety testing placed Ember below high-risk thresholds for cybersecurity, chemical and biological weapons, and loss of control.
